Re: [PATCH V4 1/3] blk-mq: move actual issue into one helper

[Mike Snitzer <snitzer@redhat.com>]

On Mon, Jan 15 2018 at 12:29pm -0500,
Jens Axboe <axboe@kernel.dk> wrote:

> On 1/15/18 9:58 AM, Ming Lei wrote:
> > No functional change, just to clean up code a bit, so that the following
> > change of using direct issue for blk_mq_request_bypass_insert() which is
> > needed by DM can be easier to do.
> > 
> > Signed-off-by: Ming Lei <ming.lei@redhat.com>
> > ---
> >  block/blk-mq.c | 39 +++++++++++++++++++++++++++------------
> >  1 file changed, 27 insertions(+), 12 deletions(-)
> > 
> > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > index edb1291a42c5..bf8d6651f40e 100644
> > --- a/block/blk-mq.c
> > +++ b/block/blk-mq.c
> > @@ -1696,15 +1696,37 @@ static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
> >  	return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
> >  }
> >  
> > -static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
> > -					struct request *rq,
> > -					blk_qc_t *cookie)
> > +static blk_status_t __blk_mq_issue_req(struct blk_mq_hw_ctx *hctx,
> > +				       struct request *rq,
> > +				       blk_qc_t *new_cookie)
> >  {
> > +	blk_status_t ret;
> >  	struct request_queue *q = rq->q;
> >  	struct blk_mq_queue_data bd = {
> >  		.rq = rq,
> >  		.last = true,
> >  	};
> > +
> > +	if (!blk_mq_get_driver_tag(rq, NULL, false))
> > +		return BLK_STS_AGAIN;
> > +
> > +	if (!blk_mq_get_dispatch_budget(hctx)) {
> > +		blk_mq_put_driver_tag(rq);
> > +		return BLK_STS_AGAIN;
> > +	}
> > +
> > +	*new_cookie = request_to_qc_t(hctx, rq);
> > +
> > +	ret = q->mq_ops->queue_rq(hctx, &bd);
> > +
> > +	return ret;
> 
> 	return q->mq_ops->queue_rq(hctx, &bd);
> 
> and kill 'ret', it's not used.

Yeap, good point.

> But more importantly, who puts the
> driver tag and the budget if we get != OK for ->queue_rq()?

__blk_mq_try_issue_directly() processes the returned value same as before
this patch.  Means this patch isn't making any functional change:
If BLK_STS_RESOURCE: __blk_mq_requeue_request() is called.
__blk_mq_requeue_request() will blk_mq_put_driver_tag().
Otherwise, all other errors result in blk_mq_end_request(rq, ret);

So ignoring this patch, are you concerned that:
1) Does blk_mq_end_request() put both?  Looks like blk_mq_free_request()
handles rq->tag != -1 but why not have it use __blk_mq_put_driver_tag()?
I'm not seeing where the budget is put from blk_mq_end_request()...

2) Nothing seems to be putting the budget in
__blk_mq_try_issue_directly()'s BLK_STS_RESOURCE error path?  I share
your concern now (for drivers who set {get,put}_budget in mq_ops).
Should __blk_mq_requeue_request() be updated to also
blk_mq_put_dispatch_budget()?

Mike